%  This is a LaTex file.

%  Homework for the course "Computational Methods in Finance", 
%  Fall semester, 1997, Jonathan Goodman.


%   A latex format for making homework assignments.


\documentstyle[12pt]{article}

%          The page format, somewhat wider and taller page than in art12.sty.

\topmargin -0.1in \headsep 0in \textheight 8.9in \footskip 0.6in
\oddsidemargin 0in  \evensidemargin 0in  \textwidth 6.5in



\begin{document}


%          Definitions of commonly used symbols.



%          The title and header.

\noindent
{\scriptsize Computational Methods in Finance, Fall 1997} \hfill 

\begin{center}
\large
Assignment 6.   
\normalsize
\end{center}

\noindent
Given  December 10, due December 31.
\vspace{.3in}

%           The questions!



\noindent
{\bf Objective:} To explore numerical optimization and statistical estimation.

\vspace{.5cm}

A very simple ARCH/GARCH type model is
\begin{equation}
\sigma_{k+1}^2 = \alpha \sigma_k^2 + \beta Z_k^2 \;\; ,
\end{equation}
\begin{equation}
X_{k+1} = X_k + \gamma \sigma_k Z_k \;\; ,
\end{equation}
where the $Z_k$ are independent standard normal ransom variables.  This is
a different approach to stochastic volatility that allows for bursts of
high volatility.  

First we want to construct maximum likelihood estimates of the parameters
$\alpha$, $\beta$, and $\gamma$.  Assume that $X_0=0$ and 
$\sigma_0 = 1$.  For given values of the parameters, the probability
density for a specific sequence $\vec{x} = (x_1,\ldots,x_n)$ is
\begin{equation}
f(\vec{x},\alpha,\beta,\gamma) = \frac{1}
       {(2\pi)^{n/2} \sigma_0\sigma_1\cdots\sigma_{n-1} }
   \exp\left( -\sum_{k=0}^{n-1}\frac{(x_{k+1} - x_k)^2}{2\gamma^2 \sigma_k^2}
                              \right)
\end{equation}
The numbers $\sigma_k$ in (3) can be computed, given $\vec{x}$, from
(1) and $\sigma_0 = 1$.  This is how $f$ comes to depend on the 
parameter $\beta$.
The maximum likelihood estimates $\hat{\alpha}(\vec{x})$, 
 $\hat{\beta}(\vec{x})$, and $\hat{\gamma}(\vec{x})$,
are found by maximizing $f$ over $\alpha$, $\beta$, and $\gamma$.

The simplest way to do this three dimensional optimization problem is
to choose search directions using the gradient method and then use
bisection search.  This requires you to compute first derivatives of
$f$ with respect to the parameters.  Make sure you program is modular: 
you should test the line search part separately, for example.
It will probably work better if you optimize $\log(f)$ instead,
the answer will be the same but the objective function will be less 
wild.  Are there local maxima that are not the global maximum?

Second, we will make ad-hoc error bars for the estimated parameter
values by a version of the jackknife process.  Produce artificial data
$\vec{x}_1$, $\ldots$, $\vec{x}_M$ by running the model (1), (2) using
the estimated parameter values.  Then use the optimization program 
on these artificial data sets to compute $\hat{\alpha}(\vec{x}_l)$
for $l=1$, $\ldots$, $l_M$.  The interval that includes $95\%$ of these
artificial $\alpha$ estimates is a poor man's $95\%$ confidence
interval for $\hat{\alpha}$.  If your computer is old or your optimization
algorithm is slow, you will not be able to take large $M$.

The data set for this problem, consisting of 20 $x$ values,
is posted on the course web site.





\end{document}